TOYON RESEARCH CORPORATION — Department of Defense SBIR Phase I: AF222-0005

Description

Track classification through feature measurements is a crucial component of any ISR system, especially those ultimately focused on target engagement. Classification algorithms typically rely on training data that consists of labeled features. The algorithm then learns the class-conditional likelihood functions for the feature. In a multi-sensor classification setting, each sensor system has been trained on its own measurement data and generates class-conditional feature likelihoods that are fused with other sensor data. This situation is more prevalent when considering a swarm of assets (e.g., UAVs or munitions) that share feature data to improve the classification of objects in the scene. There are numerous effects that cause deviations from the statistical models used in training and measurement likelihood generation. Unless these are addressed, incorrect classification will result. On this effort, Toyon will research and incorporate techniques to improve the robustness of the classification algorithms to model deviations. Robustness techniques will be applied at both the sensor processing level and the multi-sensor fusion level. The robust algorithms will be trained, tested, and evaluated on a realistic data set to determine their accuracy and efficiency when compared to the unmodified, standard classification algorithms.